Codestral

Codestral is Mistral AI's specialised family of code-generation models, distinct from its general-purpose line. The current flagship version (Codestral 25.08) sits at 22 billion parameters with a 256K-token context window and supports over 80 programming languages including Python, Java, c++, JavaScript, and Bash. It powers fill-in-the-middle (FIM) completion, function generation, and test writing at production latency, and forms the foundation of Mistral's integrated coding stack alongside Devstral 2 and the Vibe CLI. Mistral ships Codestral alongside its broader Le Chat Enterprise platform .

Released initially in May 2024 as Mistral's first code-specific model, Codestral was retrained and upgraded through versions 25.01 and 25.08, each generation cutting latency and improving benchmark scores. Codestral 25.01 achieves 86.6% on HumanEval, placing it among the top-tier open-weight coding models globally. An accompanying Codestral Embed model provides code-specific retrieval embeddings that Mistral claims outperform OpenAI and Cohere on code-retrieval benchmarks. The 25.08 release rebranded and extended the stack into a full enterprise coding suite.

For European tech-sovereignty advocates, Codestral represents a credible EU-origin alternative to GitHub Copilot (Microsoft/OpenAI, US) and Amazon CodeWhisperer. Its open-weight design allows on-premises deployment in air-gapped environments, a critical requirement for defence, finance, and public-sector customers bound by data-residency rules. Mistral's Paris headquarters and EU entity structure give it regulatory standing that US competitors lack under the EU AI Act.